Reproduce “GO-Diff: Data-free and amortized global structure optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nikolaj Rønne, Tejs Vegge, Arghya Bhowmik
The GO-Diff framework facilitates global structure optimization by employing a data-free diffusion model that learns to directly sample low-energy atomic configurations from an energy function. This approach efficiently discovers thermodynamically stable structures and enables amortized optimization through model transfer.
Key insights
- A diffusion model can effectively perform global optimization by learning directly from an energy function through a Boltzmann-weighted score-matching objective, eliminating the need for pre-existing datasets of low-energy structures.
- An annealing strategy for the Boltzmann temperature, coupled with the stochasticity of diffusion, allows for an effective balance between exploring the broader PES and exploiting specific low-energy basins, overcoming local minima traps.
- The 'generative knowledge' encoded in a pre-trained diffusion model can be successfully transferred between related atomistic systems, enabling significant computational savings and amortized optimization for new tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.